CORE GUIDE
Instruction conflict
Instruction conflict occurs when multiple valid-looking sources imply incompatible actions; reliable systems detect the conflict explicitly instead of letting the model improvise a compromise.
Mental model
A conflict is not merely confusing text. It is a decision point where two instructions cannot both be satisfied under the same state, so the system needs precedence, clarification or a safe refusal path.
Why it matters
Multi-source AI applications combine system rules, developer workflows, users, retrieved policies and tool feedback. Conflicts are therefore normal, not exceptional. If they are not represented as a first-class state, the model may choose whichever wording appears most salient, producing inconsistent behavior that is hard to debug or audit.
01
Detect, classify and resolve conflicts
Define which sources can issue instructions, then check whether their requested actions are jointly satisfiable. Classify conflicts as authority, policy ambiguity, stale state or user-goal tension. Apply the appropriate resolver: precedence for authority conflicts, clarification for ambiguity, refresh for stale evidence, and refusal or escalation when no safe action satisfies the constraints.
02
Example: user goal and compliance rule diverge
A user asks an assistant to email a complete customer record to a partner to speed up support. The user goal is legitimate, but the application policy prohibits exporting sensitive fields. The system should identify the incompatibility, preserve the user's underlying goal, and offer an allowed workflow rather than partially following both instructions.
Common failure modes
- Hiding a real conflict by asking the model to 'use best judgment'.
- Treating every conflict as a security attack when some are ordinary policy ambiguity.
- Resolving one turn without recording the underlying conflicting rule or stale state.
Engineering heuristics
- Make conflict detection observable in traces or structured state.
- Choose resolution strategy based on conflict type rather than one generic prompt.
- Preserve the user's underlying goal when refusing a conflicting requested action.
Takeaways
- 01Instruction conflicts are normal in multi-source systems.
- 02A conflict should trigger an explicit resolver, not model improvisation.
- 03Good resolution separates the user's goal from an unsafe or impossible requested action.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.